Brian Dessent wrote:
"Sebastien M." wrote:
The loop which call "empty_ptr()" is four times slower with 02.
Is there a logical reason for this ?
The only significant difference between the two seems to be that with
-O2 gcc chooses esi as the register to hold the function pointer.
-O0:
time = clock();
40136c: e8 ef 02 00 00 call 401660 <_clock>
401371: 89 45 f8 mov %eax,0xfffffff8(%ebp)
for(int i=0;i<numtests;i++) empty_ptr(x);
401374: c7 45 e4 00 00 00 00 movl $0x0,0xffffffe4(%ebp)
40137b: 8b 45 e4 mov 0xffffffe4(%ebp),%eax
40137e: 3b 45 f4 cmp 0xfffffff4(%ebp),%eax
401381: 7d 14 jge 401397 <_main+0xaf>
401383: dd 45 e8 fldl 0xffffffe8(%ebp)
401386: dd 1c 24 fstpl (%esp)
401389: 8b 45 fc mov 0xfffffffc(%ebp),%eax
40138c: ff d0 call *%eax
40138e: dd d8 fstp %st(0)
401390: 8d 45 e4 lea 0xffffffe4(%ebp),%eax
401393: ff 00 incl (%eax)
401395: eb e4 jmp 40137b <_main+0x93>
printf("Time : empty_ptr = %ld\n", clock() - time);
401397: e8 c4 02 00 00 call 401660 <_clock>
40139c: 2b 45 f8 sub 0xfffffff8(%ebp),%eax
40139f: 89 44 24 04 mov %eax,0x4(%esp)
4013a3: c7 04 24 14 30 40 00 movl $0x403014,(%esp)
4013aa: e8 a1 02 00 00 call 401650 <_printf>
-O2:
time = clock();
40133f: e8 ec 02 00 00 call 401630 <_clock>
401344: 89 c7 mov %eax,%edi
401346: 8d 76 00 lea 0x0(%esi),%esi
401349: 8d bc 27 00 00 00 00 lea 0x0(%edi),%edi
for(int i=0;i<numtests;i++) empty_ptr(x);
401350: c7 04 24 00 00 00 00 movl $0x0,(%esp)
401357: b8 00 00 08 40 mov $0x40080000,%eax
40135c: 89 44 24 04 mov %eax,0x4(%esp)
401360: ff d6 call *%esi
401362: dd d8 fstp %st(0)
401364: 4b dec %ebx
401365: 79 e9 jns 401350 <_main+0x60>
printf("Time : empty_ptr = %ld\n", clock() - time);
401367: e8 c4 02 00 00 call 401630 <_clock>
40136c: c7 04 24 14 30 40 00 movl $0x403014,(%esp)
401373: 29 f8 sub %edi,%eax
401375: 89 44 24 04 mov %eax,0x4(%esp)
401379: e8 a2 02 00 00 call 401620 <_printf>
So it could just be a bad decision by the register allocator. I don't
know what the performance difference between call *%eax and call *%esi
is.
Also note that gcc 4.3 with -O2 is smart enough to recognise that
neither of these loops do anything and remove them both entirely:
I don't have sufficient memory to recall whether the gcc version,
hardware, or compile flags information was in the original post. If you
are running on a machine where the sequence dec ebx; jns is subject to a
partial flags stall, that might be an explanation. If you used an up to
date gcc (evidently not), and wrote a loop which was not trivial to
remove (evidently not), and got such code when specifying -march for a
machine which doesn't like that sequence, this would be a bug for which
you should file a PR. Note that dec, rather than add -1, has been a
standard way to write code which might be OK on AMD but dead slow on
Intel, since 6 years ago.
